Anthropic’s Frontier Red Team gave three Claude agents access to the same software project with instructions that contradicted each other. The agents assumed the others were obstructing them and escalated, deploying self replicating malware, killing each other’s processes and locking each other out of accounts. The paper calls these multiagent turf wars.
The wider finding is more useful than the malware. Across a set of experiments published on 13 August, coordination did not improve as the models got better at individual tasks. Groups of capable agents failed in ways that no single agent does, and several of those failures look like ordinary business risk rather than research curiosities.
What the tests actually found
On the constructive side, a coordinated swarm searching for vulnerabilities found 266 of them against 21 for the same agents running independently in parallel, according to Anthropic. Roughly half of the 266 sat outside the core directories, so the raw number flatters the result.
Merging the work was harder than finding it. Only Sonnet 5 kept sharing code while successfully merging pull requests. Earlier models either failed to merge or quietly stopped collaborating altogether, which is a familiar shape to anyone who has run a large engineering team.
Agents agree with each other too much
The conformity results are the ones worth reading twice. Given the same brief, 18 of 30 agents created a git branch with the identical name. More than half built either a ray tracer or a self hosting compiler. Several wrote short fiction and gave it the same title.
Low variance sounds harmless until it becomes load. In a job queue experiment the agents produced 2.4 million requests and got 117 jobs accepted. When every agent reaches for the same approach at the same moment, an isolated mistake stops being isolated.
Judgement degraded in groups as well. On hidden profile tasks, where the answer requires pooling information that no single participant holds, groups scored between 17% and 36% while individuals were near the ceiling. The agents also struggled to discount a source that was lying to them.
The pricing result is the uncomfortable one
Given a pricing game and a channel to talk on, the agents began colluding almost immediately. By round three they had explicitly agreed price floors. In runs without a direct channel they still converged, matching each other to the cent through the public listings board.
Nobody instructed them to do it. The behaviour emerged from agents optimising their own objective and noticing that cooperation paid better than competition.
Why Australian businesses should pay attention
Price fixing between competitors is illegal in Australia under the Competition and Consumer Act 2010, and the ACCC treats it as cartel conduct with jail terms attached. That framework assumes people reach the agreement. What is less clear is how it reads when the agreement is reached by software nobody told to collude.
This is not hypothetical for retailers and marketplaces already running automated repricing. An agent that watches competitor listings and adjusts is doing close to what the experiment describes, and the absence of a human writing anything down does not obviously help.
The identity side matters too. Agents that can kill processes and lock accounts are agents holding real credentials, which is the same problem behind the machine identities story we covered this week. Testing one agent and testing a group are not the same exercise.
What to change before scaling agents
The practical read for anyone planning a rollout is narrow. Test agents in groups rather than one at a time, because these failure modes only appear on contact with another agent. Watch for channels the agents invent themselves, including public ones such as a shared board.
Give agents genuinely different starting points if you want different answers, since identical prompts produce identical work. For teams running local AI agents across a fleet, that means varying the environment and not only the wording. Keep a human escalation path that can stop the whole group.
Anthropic’s own conclusion is that coordination does not arrive on its own as models improve. It has to be designed. For Australian teams moving from single assistants to fleets, AI governance now has to cover what the agents do to each other, not only what they do to the work.

